Retrospective use of breathing motion compensation technology (MCT) enhances vessel detection software performance

Abstract

Purpose:

Cone-Beam CT (CBCT) with planning software is used in intra-arterial liver-directed therapies. Software accuracy relies on high CBCT image quality, which can be impaired by breathing motion. We assessed the impact of a specific MCT on software performance for procedure planning and navigation.

Materials and Methods:

Institutional Review Board (IRB) approved retrospective evaluation of liver-directed therapies from July 2015 to April 2018 was performed. CBCTs with at least one well-defined tumor and noticeable breathing motion were included. Each CBCT was reconstructed with and without breathing MCT (Motion Freeze, GE Healthcare). Automatic tumor-supplying vessel detection was performed on up to 4 tumors in each CBCT reconstruction (Liver ASSIST V.I., GE Healthcare). Vessel detection sensitivity and positive predictive value (PPV) were measured with and without MCT using Digital Subtracted Angiography (DSA) as reference. Preprocedural contrast enhanced CT was also utilized in some cases to rule out the possibility of extrahepatic supplying vessels.

Results:

MCT was applied retrospectively to 18 CBCTs with a total of 30 tumors. At least one supplying vessel was detected for 28/30 (93%) tumors with MCT versus 20/30 (66%) without. On the subgroup of 10 CBCTs (22 tumors, 76 feeders) in which the automatic vessel detection initially worked in both reconstructions, the average sensitivity and PPV increased from 63% (48/76) and 57% (48/84) before MCT to 83% (63/76) and 79% (63/80) after (p=0.002 and p<0.001).

Conclusion:

Breathing MCT improves planning software performance in CBCT impaired by breathing motion.

Introduction

Cone Beam Computed Tomography (CBCT) imaging is widely used in intra-arterial liver-directed therapies as it provides three-dimensional (3D) information about vascular structures and solid organ anatomy. CBCT is often performed to localize target lesions, to identify their supplying vasculature and to avoid non-target embolization [1, 2]. CBCT-based procedure planning software have been introduced a few years ago to assist interventional radiologists in detecting tumor-supplying vessels, with an increased sensitivity reported as high as 90% [3–8].

Breath hold during CBCT acquisition is necessary to ensure a proper image quality, but for some patients, it can be challenging or not simply not possible. In previous studies, significant patient motion has been reported in 20% of examinations and diaphragmatic motion in 50% of abdominal CBCTs [9]. Motion artifacts result in degraded image quality, repeated acquisitions, increased procedure time, and lower performance of tumor-supplying vessels detection software [9].

Reconstruction algorithms have been developed to improve image quality of motion impaired CBCTs. Most applied methods estimate respiratory motion models or detect patient’s motion and account for it during reconstruction [10]. Applying a newly developed MCT software, a recent study has shown improved image quality and better visualization of relevant structures for motion impaired CBCTs acquired during intraarterial liver-directed procedures [11]. In this study we investigated retrospectively the impact of the same MCT using a commercially available automatic vessel detection software to mitigate observer bias [7].

Materials and Methods

This single-center retrospective study was approved by our IRB. All patients who had a motion impaired CBCT performed during intra-arterial liver-directed therapy between July 2015 and April 2018 were reviewed. General anesthesia with breath holds during image acquisition is used in our center routinely for intraarterial liver directed therapies. Patients under general anesthesia were excluded, this represent most of our patients. Only patients who had a nonselective CBCT acquisition during hepatic arteriography, with at least one well-defined tumor were included. These patients had to have both a non-selective and selective DSA. All patients without the CBCT 2-dimensional (2D) rotational projections were excluded.

(Innova™ 4100IQ, GE Healthcare; Chicago, IL) rotating over 200° at a speed of 40°/s and an acquisition rate of 30 frames/s, obtaining a total of 150 projections in 5s.

Intra-arterial power injected DSA is routinely performed prior to CBCT acquisition to determine the optimal injection parameters. Injection locations varied from the celiac trunk to lobar hepatic arteries. CBCT images were acquired with injection of undiluted contrast agent (Omnipaque 300, GE Healthcare, Chalfont St Gilles, UK). Images were obtained with patient breath hold and injection timing delays varying from 4 to 8 s at injection rates from 1 to 5 cc/s. Contrast injection was maintained during the entire image acquisition.

Breathing motion evaluation.

Two interventional radiologists (F.R. and A.G., 6 and 10 years of experience) assessed the presence of breathing motion during the CBCT acquisition based on vessels and diaphragm motion on the CBCT 2D rotational projections. Only CBCTs with confirmed breathing motion were included in the evaluation.

Motion compensation technology (MCT).

CBCT with confirmed breathing motion were loaded to a workstation (Advantage Workstation, GE Healthcare, Chicago, IL) and reconstructed into 2 different CBCT datasets with and without MCT (Motion Freeze, GE Healthcare, Chicago, IL). This technology relies on an applied intelligence algorithm estimating the motion from the contrasted structures visible in the CBCT and fitting a respiratory motion model. The resulting motion estimation model is used during the reconstruction process to refocus the CBCT on the moving structures, thus reducing the negative effect of breathing motion. Reconstruction time is 1 s longer with MCT than without (6 s vs 5 s). The other reconstruction parameters such as matrix size, slice thickness and filters were not changed.

Supplying vessels detection.

To assess the change in CBCT image quality with vs without MCT, an automatic tumor-supplying vessel detection software (Liver ASSIST V.I., GE Healthcare, Chicago, IL) was used. For each CBCT reconstruction, hepatic arteries were semi-automatically extracted and target tumor(s) defined by one reviewer (F.R.), allowing automatic detection of supplying vessels for each identified tumor. A maximum of 4 tumors per CBCT were included. The same processing workflow was used for both datasets (with and without MCT), with the reader blinded to each dataset reconstruction method (Figure 1).

Figure 1.

First line, segment 7 tumor (A). The automatic supplying vasculature detection failed to find any vessel on the standard reconstruction (B). After motion compensation was applied, 2 supplying vessels were detected by the software (C) and confirmed by the selective DSA (D).

Second line: segment 8 tumor (E). The origin of one supplying artery was not detected correctly on the standard reconstruction (F). After motion compensation, the origin of this supplying vessel was correctly determined (G) as confirmed by the DSA (H).

Planning software performance.

Comparative analysis was conducted to evaluate performance of the planning software with and without MCT.

We compared the vascular tree automatic extraction performance on CBCTs with and without MCT. When CBCT image quality was highly deteriorated by motion artifacts, the software was unable to extract the vascular structures. We reported this scenario as vascular tree extraction failure. For these CBCTs, the number of tumor-supplying vessels visible on preprocedural contrast enhanced CT or on DSA, were accounted as false negatives.

For all CBCT reconstructions with successfully extracted vascular tree, two interventional radiologists (F.R. and A.G.) reviewed together the vessels detected by the software for each tumor. Detected vessels were classified as either true positive or false positive by consensus, using DSA as gold standard. In some cases, preprocedural contrast enhanced CT was also utilized to rule out the possibility of extrahepatic supplying vessels. False negatives, defined as supplying vessels visible on the gold standard but not detected by the software, were reported.

Virtual injection, a feature designed to help interventional radiologists determining optimal delivery point by simulating on the proximal CBCT any selective injection [12], was tested among the true positive tumor-supplying vessels and its accuracy was determined against DSA.

Sensitivity and positive predictive value (PPV) of tumor-supplying vessel automatic detection software were defined for each CBCT reconstruction. To mitigate the impact of vascular tree extraction failure, PPV was only calculated on the subgroup of CBCTs with successfully extracted vascular tree. Sensitivity was calculated on this same subgroup, as well as on the full tumor cohort. Within this subgroup, we classified them by location of the tumors; central (segments I, IV, and V) or peripheral (segments II, III, VI, VII, and VIII) and analyzed the corresponding sensitivities and PPVs.

Statistical analysis.

Sensitivity and PPV were compared by cases using Wilcoxon signed-rank test. Statistical analysis was performed using R software, and the statistical significance limit was set at p < 0.05.

Results

Patient characteristics.

87 CBCTs with 2D rotational projections available were screened, 30% (26/87) had breathing motion artifacts. Eight CBCTs were excluded because the injection was too selective or because no tumor was found. The MCT was applied retrospectively to the remaining 18 CBCTs with a total of 30 tumors. Thirteen tumors (43%) were centrally located (segments I, IV, and V) and seventeen tumors (67%) were peripherally located (segments II, III, VI, VII, and VIII) Table 1.

Table 1:

Characteristics

Variable	Value
Number of CBCTs	18
Tumors (n)	30
Average size	37 (± 17) mm
Supplying vessels (n)	94
Location	segment I (n=1), segment III (n=1), segment IV (n=3), segment V (n=9), segment VI (n=5), segment VII (n=4), segment VIII (n=7)

Using both DSA and pre-procedural CT as reference, a total of 94 supplying arteries were identified with an average of 3.2 supplying arteries/tumor.

Planning software performance

Without MCT, the software was unable to extract vessels for 8 of the CBCTs (44%) involving 8 tumors. No supplying vessel were detected for 2 other tumors despite successful extraction of the vascular tree, resulting in a detection failure of any supplying vessel for a total of 10 tumors (33%). After applying MCT, the vascular tree was extracted in all CBCTs and at least one feeding vessel was detected for 28/30 tumors (93%), as illustrated in figure 1.

Forty-eight true positives, representing 51% of total real supplying vessels, and 38 false positives supplying vessels were identified without MCT. With MCT, 76 true positives, i.e. 81% of total real supplying vessels, and 20 false positives supplying vessels were identified.

For all CBCTs, the sensitivity of vessel detection was significantly higher with motion correction than without (81% versus 51% respectively, p <0.006).

The virtual injection was described as accurate in 64/76 true positive supplying vessels (84%) with MCT versus 21/48 (44%) without MCT. Comparison of software performances with and without MCT is summarized in table 2 and figure 2.

Table 2:

Planning software performance summary

	MCT OFF	MCT ON		Subgroup analysis (10 CBCTs)
Vascular tree successfully extracted (n=18)	10	18			MC* OFF	MC* ON	p value
True positive tumor-supplying vessels (n=94)	48	76	All tumors (n=22)	Sensitivity	63% (48/76)	83% (63/76)	0.002
				PPV	57% (48/84)	79% (63/80)	0.001
False positive tumor-supplying vessels	38	20	Central lesions (n=10)	Sensitivity	76% (28/37)	89% (33/37)	0.060
Sensitivity	51% (48/94)	81% (76/94)		PPV	61% (28/46)	79% (33/42)	0.022
Virtual injection prediction accuracy among true positive tumor-supplying vessels	21/48	64/76	Peripheral lesions (n=12)	Sensitivity	51% (20/39)	77% (30/39)	0.015
				PPV	53% (20/38)	79% (30/38)	0.006

Figure 2 -. Motion compensation impact on automatic tumor supplying vessels performance.

Analysis of the software performance with and without motion compensation on 4 criteria. Criterion A is expressed as ratio between tumors for which a vascular tree was extracted by the software and the total number of tumors as explained in material and methods. The accuracy of virtual injection has been tested among true positive supplying vessels and is reported in D.

Subgroup analysis

Impact of MCT was evaluated in CBCTs where the vascular extraction had not failed in one or the other reconstruction method. For this subgroup of 10 CBCTs (22 tumors supplied by 76 arteries), sensitivity and PPV of tumor-supplying vessel automatic detection were calculated and found significantly higher after motion correction (83% and 79% versus 63% and 57% respectively, p <.002).

The improvement in sensitivity and PPV was not uniform across the tumor locations. While there was no statistically significant difference in detection sensitivity without versus with MCT for the 10 centrally located tumors (76% versus 89% respectively, p=0.06), there was a significant improvement for the 12 peripherally located tumors (51% vs 77%, p=0.015). The PPV was significantly increased in both subgroups with MCT, from 61% to 79% for centrally located tumors (p=0.022) and from 53% to 79% for peripheral tumors (p=0.006).

Discussion

In this study the retrospective use of MCT demonstrated to have a positive impact on automatic vessel detection software performance in CBCTs corrupted by breathing motion. MCT resulted in a significant improvement in detection of tumor-supplying vessels with a sensitivity of 81% (for the whole cohort) and positive predictive value of 79% (for the subgroup analysis).

Previous studies using tumor-supplying vessel automatic detection software on CBCTs free of respiratory motion artifacts reported sensitivities of 80 – 97% and positive predictive values of 83 – 99% [4–8], which is similar to our results. Undetected tumor-supplying vessels are often small arteries (<1mm) difficult to identify and generally requiring selective DSAs [13].

Improvement after MCT was better for peripherally located tumors, probably because they are more susceptible to motion artifacts. Even if the MCT did not significantly increase the software sensitivity for central tumors, it still minimized the number of false positive tumor-supplying vessels.

This MCT software was previously evaluated by Burgio et al. based on a visual score. In this study, we believe the method used is less subjective and more reproducible, thanks to the utilization of an automatic tumor supplying vessel software as indicator of CBCT image quality. Their conclusions nevertheless corroborate the results of the current study, both in terms of improvement in image quality and potential decrease of CBCT reacquisition [11]. Maximum intensity projection (MIP) images improve overall vessels visibility and visual detection of tumor-supplying vessels [14], with no impact on automated software performance.

More sophisticated MCT methods have been developed to compensate respiratory motion of CBCT, for both intraprocedural imaging and radiotherapy planning. Some methods are based on non-rigid registration of a 3D vascular tree to 2D contrast enhanced projections [15], others are using iterative techniques to fit a 3D respiratory model on CBCTs performed in free breathing [10]. Both methods have shown a significant improvement in image quality, but they are time-consuming and not compatible with clinical routine.

There are relevant limitations of this study. First, the evaluation of breathing motion on the CBCT 2D rotational views can be challenging, as it is difficult to distinguish between motion related to cardiac activity, patient movements, and respiratory cycle. The gold standard for tumor-supplying vessels was DSAs obtained during procedures. Preprocedural contrast enhanced CT was also used in some cases to rule out the possibility of extrahepatic supplying vessels. This leaves room for missed arteries by virtue of operator vessel selection during the procedure and by the selection of DSA images. Manual definition of the tumor ROI between the different series (with and without MCT) could be responsible for detection differences, even though all efforts were made to keep all parameters constant. Common concern for MCT methods is inaccurate respiratory model fitting, due to metal artifact or extreme diaphragm motion for example, which could lead to image quality degradation or vascular structure distortion. Finally, this is a retrospective study with a small number of patients.

In conclusion, MCT improves planning software performance for CBCTs impaired by breathing motion artifacts. MCT increases tumor-supplying vessels automatic detection sensitivity and PPV. This technology can potentially reduce the need for repeated acquisitions and therefore additional exposure to radiation and contrast, ideally this should be investigated in a prospective study.